unflatten_op

→TensorImplPtr

int unflatten_op(const int & a, int dim, const int & sizes)

source edit

Python wrappers

pylucid.ops.unflattenfunction

Inverse of flatten — split dim into sizes.

Composite over reshape_op. The product of sizes must equal a.shape[dim]; the surrounding axes are preserved verbatim. Pure view operation when the input is contiguous; gradient flows through ReshapeBackward.

Math

Let a have shape $(s_0, \ldots, s_{d-1}, s_d, s_{d+1}, \ldots)$ and sizes $= (m_0, \ldots, m_{k-1})$ with $\prod_i m_i = s_d$ . Then the output has shape

(s_0, \ldots, s_{d-1}, m_0, \ldots, m_{k-1}, s_{d+1}, \ldots)

See Also

reshape_op — the underlying primitive operates on the full shape.

Parameters

aTensorImplPtr

Source tensor of rank

\ge 1

dimint

Axis to expand. Negative values wrap modulo a.ndim.

sizesShape

Replacement extents. Their product must equal a.shape[dim].

Returns

TensorImplPtr

Reshaped tensor whose total numel matches a.

Raises

Failure

If a is null or

\prod \mathrm{sizes} \ne \mathrm{a.shape}[\mathrm{dim}]

IndexError

If dim is out of range.